Sol sold some of its Dell shares
12 August 2026
Dell's share of the portfolio5.2%
New target3.5%
After Dell rallied, Sol used the risk tool to check a smaller holding. Shares were sold the next morning.
No evidence of forecasting skill in four of five models
Five frontier models ran a long/short book for two months inside a terminal rebuilt from the public record. Only one showed a usable forecast signal, and its book still finished level with the index.
- Models
- 5
- Decision dates
- 43
- Stocks
- 498
- Window
- 2 mo3 Aug to 1 Oct 2026
- SPY over the window
- +1.08%
- Beat the index
- 2 of 5
Three of five lost to the index
Sol · OpenAI
Opus · Anthropic
Fable · Anthropic
Astra · Google
Grok · xAI
SPY
Only one model showed a forecasting signal
Each model forecast five-day returns relative to SPY across 498 stocks, then traded on its own calls. Only Fable's signal cleared the usual 0.05 floor, and its book still finished level with the index.
Forecast accuracy (IC)
Rank correlation of forecast and realized five-day returns · 38 resolved dates · Dashed: 0.05, the usual floor for a usable signal
Return vs SPY
Percentage points over the index · Three of five below zero
Information ratio
Beta-adjusted return per unit of risk · Below zero is worse than SPY
Largest drawdown
Peak-to-trough loss · Shorter is better
| Model | Forecast IC | Return | vs SPY | Info. ratio | Max. drawdown |
|---|---|---|---|---|---|
| Sol | 0.026 | +4.56% | +3.47 pp | 1.61 | −3.13% |
| Opus | 0.029 | +2.11% | +1.03 pp | 0.96 | −2.19% |
| Fable | 0.073 | +1.08% | −0.004 pp | 0.40 | −5.03% |
| Astra | 0.014 | +0.39% | −0.70 pp | −0.37 | −2.83% |
| Grok | 0.010 | +0.52% | −0.57 pp | −1.30 | −4.69% |
Ordered by information ratio. SPY returned +1.08%. Metrics are computed per run, then averaged by model.
Forecast accuracy (IC)
Rank correlation between predicted and realized five-day stock returns relative to SPY. Zero means no ranking signal. A common rule of thumb treats 0.05 as the floor for a usable signal.
Information ratio
Annualized beta-adjusted daily return divided by its volatility, with the benchmark's downside guard.
Forecast IC covers 38 resolved dates; five remain pending.
Returns include the Vylor shares owed on Corteva shorts, valued at the October 1 close.
How they went wrong
Four documented errors. Each was identifiable from the data the model had in front of it at the time.
- Grok called a guidance cut a raise. Tractor Supply cut its full-year earnings guidance from $2.13–2.23 to $1.78–1.88 per share. Relying on an incorrect feed label, Grok wrote "FY EPS and margin raised" and closed its 2% short position on July 23.
- Tool failures became holds. A missing-usage error caused Grok to hold its book on 7 July instead of deciding. Missing-document errors also occurred during the run. In both cases the model carried on without flagging the gap.
- Fable sold on a duplicate earnings date. SanDisk had reported its June-quarter results on August 5. A duplicate calendar entry listed August 24; Fable wrote "reports 08-24 within 9-day rule" and sold its entire position on August 14.
- Sol read a smaller loss as worse earnings. Gilead narrowed its projected adjusted loss from $0.65–1.05 to $0.30–0.65 per share. Following an incorrect feed label, Sol wrote "GAAP and adjusted EPS guided LOWER" and cut its five-day return forecast by 1.60 percentage points. Only GAAP guidance worsened. No Gilead trade followed.
These errors were identifiable at the time. They do not establish that the opposite decisions would have made money.
The two decisions that worked
Opus sold its entire MGM holding
24 September 2026
Searched MGM news: nothing found
Checked a shareholder filing: the proposed takeover was called off
Decision: sell all MGM shares
The sale went through the next morning.
The terminal we built
Data rebuilt through time from the public record.
How we rebuilt a terminal from the public record
Public-domain records in the store, by dataset (log scale)
As published by the SEC, Fed, Treasury and EIA
Extracted by us from 8-K text, each row quoting its source
Counts as of 8 October 2026. Prices, index membership and news come from open or unofficial feeds and are not shown here.
Rebuilt from headlines and filings
The sell-side layer a terminal normally licenses was reconstructed from what companies and the wire say in public. Three real records from the store:
1. An earnings date, from a wire headline
“Boeing to Release Second Quarter Results on July 28”
prnewswire.com · seen on the wire 1 Jul 2026
- SYMBOL
- BA
- DATE
- 2026-07-28
- SOURCE
- announced · visible 27 days ahead
- RIGHTS
- headline only, GDELT terms
59 of 59 headline-sourced dates exact
2. An analyst action, from a wire headline
“Tigress Financial Lifts NVIDIA (NVDA) Price Target, Keeps Strong Buy Rating”
insidermonkey.com · seen on the wire 1 Jun 2026
- SYMBOL
- NVDA · named in the headline
- KIND
- analyst · price target raised
- RATING
- Strong Buy, reiterated
- REPORT
- never fetched, never stored
26,261 analyst actions from headlines alone
3. Guidance, from an 8-K exhibit
“Abbott projects full-year 2025 adjusted diluted earnings per share of $5.05 to $5.25”
accepted by EDGAR 16 Apr 2025 07:37 ET
- SYMBOL
- ABT
- MEASURE
- adjusted EPS · FY2025 · non-GAAP
- RANGE
- $5.05 to $5.25
- KNOWABLE
- acceptance + 60 s, quote required
9,027 guidance lines, each quoting its release
1.3Mwire items since Jan 2025, stamped at first-seen time
533 / 545event-days with a forward calendar rebuilt for that day
2,220consensus records recovered from Wayback captures, labelled retrospective
514 / 533release timings read correctly from the filing text
Evaluation notes
Read before quoting. The window is short and the sample is small.
3 Aug–1 Oct 2026 · 498 stocks · 43 decision dates
Terminal-only research. Decisions fill at the next open; final-day orders remain unfilled.
We chose a short, recent window to keep evaluation outcomes after the models' stated pretraining cutoffs. Noise is substantial over just 43 decision dates: returns and rankings are not evidence of persistent investment skill. Starting portfolios differ, and individual information-ratio confidence intervals include zero.
Original ledger replay matched recorded equity. Report valuations include the spin-off correction; historical trades and fees are retained. A missing-usage error caused Grok to hold on 7 July. Missing-document errors also occurred.
Product: daily data, no live execution; incomplete document coverage. Production reliability and data rights remain work in progress.
Scoring versions: book-v4, events-v3, universe-v2. All returns are simulated.
Run your model on it
We will run any model you name through the same terminal and window, and send you every decision, forecast and ledger. Contact us for access to the codebase.
